GF Securities: NVIDIA Corporation (NVDA.US) new platform enhances the competitiveness of Agent applications, AI inference drives storage cycle continuous growth.

GF SEC released a research report stating that NVIDIA Corporation (NVDA.US) showcased Vera Rubin POD at GTC, focusing on enhancing the competitiveness of Agent applications in cluster computing power and inference computing power product lines. At the moment of AI's advancement, model innovation and CAPEX foundation are crucial for the collaborative development of the AI industry chain; AI inference-driven storage cycles continue to rise, with expanded production and upgrades working together. It is recommended to focus on core beneficiaries of the industry chain. Key points from GF SEC: NVIDIA Corporation releases Vera Rubin POD platform According to NVIDIA Corporation's official website, on March 16, 2026, NVIDIA Corporation showcased Vera Rubin POD at GTC, including 5 new rack-level systems specifically designed for Agentic AI workloads. Due to the higher demands of Agentic workloads for high throughput, extremely low latency inference, dense CPU sandbox, and massive context memory storage, NVIDIA Corporation focused on enhancing the competitiveness of Agent applications in cluster computing power and inference computing power product lines. Vera Rubin POD is mainly divided into two types of racks: (1) MGXNVL rack: Vera Rubin NVL72, the rack internally connected via NVLink, is responsible for core GPU computing tasks; (2) MGXETL rack: includes Groq3 LPX rack, Vera CPU rack, BlueField-4 STX storage rack, and Spectrum-6 SPX network rack, all interconnected directly through SpectrumX Ethernet or Groq3 LPU chips. According to the official website diagram, a Vera Rubin 1152 SuperPOD consists of 16 Vera Rubin NVL72 racks, 2 Vera CPU racks, 10 Groq 3 LPX racks, 2 BlueField-4 STX storage racks, and 10 Spectrum-6 SPX network racks, reflecting its heterogeneous collaborative system architecture built around Agentic AI. Groq3 LPX rack used for acceleration of decoding. The Groq3 LPX rack integrates 256 LPU processors, equipped with 128 GB on-chip SRAM and 640TB/s bandwidth. In the combination architecture of Vera Rubin NVL72 and LPX, the GPU is mainly responsible for Prefill and Attention calculations in the Decode phase, while the LPU accelerates the FFN calculations in the Decode phase, speeding up the decoding process of each output token per layer, and collaborating with Vera Rubin racks through custom Spectrum-X interconnect. As disclosed on the NVIDIA Corporation's official website, under conditions of 400 TPS per user, the combination of Vera Rubin NVL72 and LPX can achieve up to 35 times higher TPS increase per megawatt compared to NVIDIA GB200 NVL72, improving the overall system performance and being more suitable for low latency and highly interactive Agent application scenarios. Vera CPU rack supports RL/Agent sandbox environments The Vera CPU rack integrates 256 Vera CPUs, adopts a high-density liquid cooling design, and can support over 22500 concurrent reinforcement learning (RL) or agent sandbox environments in a single rack, used for testing, executing, and verifying the output results of Vera Rubin NVL72 and LPX. Risk Warning AI industry development and demand lower than expected; AI server shipments lower than expected, domestic manufacturers' technological and product progress lower than expected.